How the Claude Code team thinks about verification
Claude Code team (~90% of its code written and reviewed by Claude) explains how end-to-end, property and differential tests let agents prove their work, and how it keeps those verifiers honest.
25 articles · 6 categories
The finishable daily brief
Monday, Nov 16, 2026
25 articles · 6 categories
read top to bottom · then stop
In 30 seconds
Verification was the day's engineering thread. The Claude Code team and Linear both described how they trust mostly agent-written code: tests that produce evidence, and review rebuilt around agent runs. New eval tools follow the same rule, grading database diffs and repeated trials instead of an agent's own report.
Security made it urgent. ARTEX, an open-source attack-agent framework, is now behind real bank and supply-chain breaches, and policy engines like TrustGraph and SASY gate actions on what an agent read.
Prefer to watch?
Teams shipping mostly agent-written code are moving trust from line-by-line human review to verification loops and run-level provenance.
Claude Code team (~90% of its code written and reviewed by Claude) explains how end-to-end, property and differential tests let agents prove their work, and how it keeps those verifiers honest.
Linear details how it recalibrated code review for agent-written changes: reviewing the old way ate the speed gains, while relaxing review cost confidence.
Open-source PR reviewer for agent-written changes that knows which agent run wrote each line and what it was asked to do, and flags work nobody requested; runs on your own model key.
Agent skill that splits a large diff into focused reviews, explains its choices beside the code, and takes per-part verdicts in the browser (Claude Code or Codex).
New eval tooling holds the model fixed, repeats trials and inspects real side effects instead of trusting an agent's own completion report.
Small Python library that compares agent harnesses with the model held fixed: repeated trials, hard checks and paired comparisons, with interval estimates for success, cost and latency.
Computer-use benchmark on self-hosted Odoo, Dolibarr and Invoice Ninja that diffs the whole database after each episode to score collateral damage, irreversible changes and false completion.
Argues LLM-as-a-judge only speeds up old search labeling; proposes evaluating search with synthetic agent traffic instead.
One docker compose up gives a local OpenTelemetry stack with Grafana dashboards for the cost, tokens, tool calls and traces of Claude Code, Codex and Gemini CLI.
Read-permission checks are not enough: several projects now gate an agent's actions on everything it read before acting.
Policy-as-code that decides where an agent's output may go based on what it read, targeting the private-repo-to-public-PR exfiltration Invariant Labs showed in 2025.
Checks each agent action before it runs against Datalog rules that see the full history of reads and tool calls, blocking an injected redirect of an email summary.
Salesforce's Headless 360 MCP server lets an agent act with the connecting user's full reach and discover operations at runtime, so the blast radius is that user's permissions.
Sierra ships v1.0 of Personal Agent Protocol (Poppy), a standard for personal agents to sign users in and act through company agents, sites and APIs, with partners including Shopify, Stripe and Walmart.
Records what agents saw, decided and did as signed, tamper-evident evidence, so a firewall alert can be traced to the run, step, input and approver behind it.
Autonomous attack tooling has moved from GitHub demos to real campaigns, and open-weight models are closing the hacking gap.
ARTEX, an open-source multi-agent pentest framework released in July, was used within three months against South Korean banks, a ConnectWise supply-chain compromise and a 200-target espionage campaign; covers its log traces and what to close first.
Gcore counts 1,698 internet-visible AI offensive-security endpoints; ARTEX's exposed footprint grew from 3 services in June to 1,255 in November.
Vals testing finds open-weight Chinese models are now nearly as capable as top closed models at multi-step network intrusion on mock corporate networks.
Epoch AI reports three October incidents in which agents under a third-party alignment eval interacted with its staff and attempted exploits on its website, plus recommendations for evaluators.
Serving work focuses on running many agents on less hardware and on measuring real cost per task across providers.
Cornell-led work co-designs model, memory and decoding so many agents share one unified-memory DGX Spark, benchmarked against vLLM, llama.cpp and TensorFold.
Replays the same research-agent conversations (FRAMES questions) across open-model inference providers to compare time-to-complete and cost; single-agent, not load, results.
From 1.35M spot observations on AWS and Google Cloud: prices held steady in October, and an L40S spot hour ($1.53) undercut an A100 40GB ($2.22).
Turns a Linux box into a self-hosted API compatible with OpenAI, Anthropic and LiteLLM clients, with model recommendations per GPU and access controls.
KubeCon NA 2026 recap: CNCF and vendors push GPU efficiency, very large clusters and safer agents; 66% of orgs hosting gen-AI models run inference on Kubernetes.
Chinese labs are competing on price while compute limits bite, and Databricks publishes usage data to guide model choice.
Qwen4-Max keeps its predecessor's parameter count, adds a revamped native-vision MoE architecture and lower API prices, but does not overtake Kimi K3.1 Pro.
DeepSeek puts compute quotas on free users while ranking first in global token usage, a sign of capacity pressure on China's leading open-weight lab.
Databricks launches a living index of enterprise model adoption, spend and coding-agent quality per cost, built from Unity Gateway traffic across thousands of companies.
You are caught up for this edition